Clemson University Clemson University
TigerPrints TigerPrints
All Theses Theses
May 2020
Worst Case Datasets for Solving Binary Logistic Regression via Worst Case Datasets for Solving Binary Logistic Regression via
Deterministic First-Order Methods Deterministic First-Order Methods
Trevor Squires Clemson University, [email protected]
Follow this and additional works at: https://tigerprints.clemson.edu/all_theses
Recommended Citation Recommended Citation Squires, Trevor, "Worst Case Datasets for Solving Binary Logistic Regression via Deterministic First-Order Methods" (2020). All Theses. 3282. https://tigerprints.clemson.edu/all_theses/3282
This Thesis is brought to you for free and open access by the Theses at TigerPrints. It has been accepted for inclusion in All Theses by an authorized administrator of TigerPrints. For more information, please contact [email protected].
Worst Case Datasets for Solving Binary LogisticRegression via Deterministic First-Order Methods
A Thesis
Presented to
the Graduate School of
Clemson University
In Partial Fulfillment
of the Requirements for the Degree
Master of Science
Mathematical Sciences
by
Trevor Squires May 2020
Accepted by:
Dr. Yuyuan Ouyang, Committee Chair
Dr. Fei Xue
Dr. Christopher McMahan
In this thesis, we construct worst case binary logistic regression datasets for any determin-
istic first order methods. We show that our datasets require at least O(1/√ε) first-order oracle
inquires to obtain a ε−approximate solution under the assumption that the problem dimension is
sufficiently large. Using our result, on worst case datasets we conclude that existing algorithms such
as Nesterov’s Optimal Gradient Descent are optimal algorithms for solving binary logistic regression
under large scale assumptions. Our analysis combines Nemirovski’s Krylov subspace technique and
Nesterov’s construction of worst case convex quadratic programming problem instance. Our result
is the first worst case dataset constructed against all first order methods for solving binary logistic
regression, and a new worst case instance among smooth convex optimization problems.
ii
Acknowledgments
This research is supported by National Science Foundation grant DMS-1913006.
iii
Table of Contents
Title Page . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Problem of Interest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Properties of Smooth Convex Functions . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Solving the Binary Logistic Regression Problem . . . . . . . . . . . . . . . . . . . 92.1 A First Order Method for Smooth Convex Optimization . . . . . . . . . . . . . . . . 92.2 Convergence Analysis of Accelerated Gradient Descent . . . . . . . . . . . . . . . . . 10
3 Optimality of Nesterov’s Optimal Gradient Method . . . . . . . . . . . . . . . . . 153.1 Iteration Complexity Lower Bound for Smooth Convex Optimization . . . . . . . . . 16
4 Lower Complexity Bound for Binary Logistic Regression . . . . . . . . . . . . . . 244.1 Linear Span Assumption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254.2 Lower Complexity Bound for solving Binary Logistic Regression via General first
order Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
iv
Chapter 1
Introduction
1.1 Problem of Interest
In statistics, binary logistic regression is used to model the probabilities of a particular pair
of (typically mutually exclusive) characteristics. For example, binary logistic regression appears
frequently in biostatistical applications as binary responses such as health status (healthy/sick) or
survival status (alive/dead) are common. In this thesis, our main concern is solving binary logistic
regression problems. Given any data matrix A ∈ RN×n whose rows aTi represent input data and
response vector b ∈ {−1, 1}N , we assume that
P (b(i) = 1 | aTi ;x, y) =1
1 + exp(−aTi x+ y)
for some parameters x ∈ Rn, y ∈ R to be estimated. Throughout this thesis, we use b(i) to denote
the i-th component of a vector b. Noting that,
P (b(i) = −1 | aTi ;x, y) = 1− P (b(i) = 1 | aTi ;x, y) =exp(−aTi x+ y)
1 + exp(−aTi x+ y)=
1
1 + exp(aTi x+ y),
we see that the probability mass function can be written as
p(b(i) | aTi ;x, y) =1
1 + exp(−b(i)aTi x+ y). (1.1)
1
To estimate parameters x and y, we proceed by maximizing the likelihood function, or equivalently,
the log-likelihood function l : Rn × R→ R given by
l(x, y) =
N∑i=1
log
(1
1 + exp(−b(i)aTi x+ y)
)= −
N∑i=1
log(1 + exp(−b(i)aTi x+ y)
).
Since a factor of 2 does not change the minimizer, we say that the binary logistic regression problem
is an optimization problem of the form
minx∈Rn,y∈R
N∑i=1
2 log(1 + exp
(−b(i)(aTi x+ y)
)). (1.2)
For theoretical analysis, it may be more helpful to write (1.2) more succinctly. Define for any u ∈ Rk
h(u) ≡ hk(u) :=
k∑i=1
2 log(
2 cosh(u(i)
2
))=
k∑i=1
2 log(
exp(u(i)
2
)+ exp
(−u(i)
2
)).
(1.3)
Now, denoting
φA,b(x, y) := h(Ax+ y1)− bT (Ax+ y1) (1.4)
where 1N ∈ RN is a vector of 1’s, we can reformulate (1.2) as
minx∈Rn,y∈R
φA,b(x, y). (BLR)
Indeed, under the definition of h in (1.3), we have
φA,b(x, y) =
N∑i=1
2 log
(exp
(aTi x+ y
2
)+ exp
(−a
Ti x+ y
2
))− b(i)(aTi x+ y)
=
N∑i=1
2 log
(exp
(b(i)(a
Ti x+ y)
2
)+ exp
(−b(i)(a
Ti x+ y)
2
))− b(i)(aTi x+ y)
=
N∑i=1
2 log(1 + exp
(−b(i)(aTi x+ y)
)).
Therefore (BLR) is equivalent to (1.2).
2
1.2 Properties of Smooth Convex Functions
In this section, we discuss and prove a few properties of smooth convex functions that will
be helpful in solving (BLR). In particular, we prove that the log-likelihood function of the binary
logistic regression model is convex and smooth. Let us begin with a definition.
Definition 1.1. A function f : Rn → R is said to be convex if for any x, y ∈ Rn and any λ ∈ [0, 1],
f(λx+ (1− λy)) ≤ λf(x) + (1− λ)f(y). (1.5)
In addition, we say that f is concave if −f is convex.
When f is differentiable, we have an equivalent definition of convexity:
Proposition 1.1. A differentiable function f : Rn → R is convex if and only if for any x, y ∈ Rn
f(y) ≥ f(x) + 〈∇f(x), y − x〉. (1.6)
Proof. Let λ ∈ [0, 1] and suppose that f is convex, i.e. (1.5) holds. Then by rearrangement,
f(y) ≥ f(λx+ (1− λ)y)− λf(x)
1− λ
≥ f(λx+ (1− λ)y)− f(x)
1− λ+ f(x)
holds. Now apply a change of variable h = 1− λ so that we obtain
f(x) +f(1− h)x+ hy)− f(x)
h≤ f(y).
Letting h→ 0, we conclude that
〈∇f(x), y − x〉+ f(x) ≤ f(y).
Conversely, suppose that (1.6) holds. Then we have
f(x) ≥ f(λx+ (1− λ)y) + 〈∇f(λx+ (1− λ)y)〉, x− λx− (1− λ)y))
f(y) ≥ f(λx+ (1− λ)y) + 〈∇f(λx+ (1− λ)y)〉, y − λx− (1− λ)y)).
3
Let z = λx+ (1− λ)y. Applying these inequalities to λf(x) + (1− λ)y gives
λf(x) + (1− λ)f(y) ≥ λ (f(z) + (1− λ)〈∇f(z), x− y〉) + (1− λ) (f(z) + λ〈∇f(z), y − x〉)
= f(z).
Thus, f is convex.
For twice differentiable functions, we may also use the following corollary to verify convexity:
Corollary 1.2. A function f : Rn → R with ∇2f(x) � 0 for all x ∈ Rn is convex.
Proof. Let f satisfy the corollary and x, y ∈ Rn. Consider the Taylor series expansion of f(y) about
the point x.
f(y) = f(x+ y − x) = f(x) + (y − x)T∇f(x) + (y − x)T∇2f(ξ)
2(y − x)
for some ξ ∈ Rn. Since ∇2f(ξ) � 0 by assumption, it follows that
f(y) ≥ f(x) + (y − x)T∇f(x).
Thus f is convex.
Additionally, there are a few operations that preserve convexity. Here, we list two that will
be of use to us.
Proposition 1.3. Let {fi}mi=1 be a set of convex functions with fi : Rn → R for all i and {ci}mi=1
be a set of constants satisfying ci ≥ 0 for all i. Then f =∑m
i=1 cifi is a convex function.
Proof. Let x, y ∈ Rn and λ ∈ [0, 1]. Then we have
f(λx+ (1− λ)y) =
m∑i=1
cifi(λx+ (1− λ)y)
≥m∑i=1
ci (λfi(x) + (1− λ)fi(y))
= λf(x) + (1− λ)f(y).
Thus f is convex.
4
Proposition 1.4. Let g : Rn → R be a convex function, A ∈ Rn×m, and b ∈ Rn. The function
f : Rm → R defined by
f(x) = g(Ax+ b)
is convex.
Proof. Let f, g, A, and b be defined as above, fix points x, y ∈ Rm, and let λ ∈ [0, 1]. Then
f(λx+ (1− λ)y) = g(A(λx+ (1− λ)y) + b)
= g(λ(Ax+ b) + (1− λ)(Ay + b))
≥ λg(Ax+ b) + (1− λ)(Ay + b)
= λf(x) + (1− λ)f(y).
We conclude that f is convex.
In addition to convexity, we will also require some smoothness for later sections. The
following definition and proposition will prove useful.
Definition 1.2. We say that function f : Rn → R is smooth (with respect to ||·||) if it is differentiable
and ∇f is Lipschitz continuous with Lipschitz constant L > 0:
||∇f(x)−∇f(y)||∗ ≤ L ||x− y|| (1.7)
for any x, y ∈ Rn. Here, ||·|| is any norm and ||·||∗ is its dual norm defined as
||x||∗ = sup||y||≤1
〈x, y〉
with x ∈ Rn. Although this definition holds in general, for the rest of this thesis, we will use the
Euclidean norm ||·||2. Note that ||·||2 is self-dual.
Proposition 1.5. Suppose that f : Rn → R is a convex function that additionally satisfies (1.7).
Then,
|f(y)− f(x)− 〈∇f(x), y − x〉| ≤ L
2||y − x||2 (1.8)
for all x, y ∈ Rn.
5
Proof. We begin by defining an auxillary function
g(λ) := f((1− λ)x+ λy) (1.9)
for a fixed pair x, y ∈ Rn and λ ∈ [0, 1]. It follows that by chain rule,
g′(λ) = 〈∇f((1− λ)x+ λy, y − x〉.
Since g(1) = f(y) and g(0) = f(x), we may apply the fundamental theorem of calculus to obtain
g(1) = g(0) +
∫ 1
0
g′(λ)dλ
Applying the definition (1.9), we may expand this to
f(y) = f(x) + 〈∇f(x), y − x〉+
∫ 1
0
〈∇f(1− λ)x+ λy)− f(x), y − x〉dλ.
By application of Holder’s inequality, we obtain
|f(y)− f(x)− 〈∇f(x), y − x〉| =∣∣∣∣∫ 1
0
〈∇f((1− λ)x+ λy)−∇f(x), y − x〉 dλ∣∣∣∣
≤∫ 1
0
|〈∇f((1− λ)x+ λy)−∇f(x), y − x〉| dλ
≤∫ 1
0
||∇f((1− λ)x+ λy)−∇f(x)||∗ ||y − x|| dλ
≤∫ 1
0
Lλ ||y − x||2 dλ
=L
2||y − x||2 .
The following corollary is a direct result of (1.6) and (1.8).
Corollary 1.6. For any convex differentiable function f : Rn → R satisfying (1.7), we have
0 ≤ f(y)− f(x)− 〈∇f(x), y − y〉 ≤ L
2||y − x||2 (1.10)
6
for all x, y ∈ Rn.
With tools developed in this chapter, we are now ready to show that (BLR) is a smooth
convex optimization problem.
Theorem 1.7. The objective function φA,b(x, y) in (BLR) is convex and has Lipschitz continuous
gradient.
Proof. By Proposition 1.3, to show that function φA,b(x, y) = h(Ax+y1)− bT (Ax+y1) is convex, it
suffices to show that h(Ax+ y1) and −bT (Ax+ y1) are both convex. The former is the composition
of a function h and an affine transformation. By Proposition 1.4, if h is convex, then so is h(Ax+y1).
From the definition of h in (1.3), it is easily verified that ∇2h(x) is diagonal with
[∇2h(x)]ii =cosh2
(x(i)
2
)+ sinh2
(x(i)
2
)cosh2
(x(i)
2
) > 0,∀x.
For the latter term, note that −bT (Ax + y1) is the composition of an affine function and a linear
function which is verifiably convex from the definition in (1.5). Thus φA, b is a convex function.
Observe that by defining z =
xy
, D =
(A 1N
)we have
∇φA,b(z) = DT∇h(Dz) +DT b.
Continuing, let z1, z2 ∈ Rn+1 be two vectors. Then
||∇φA,b(z1)−∇φA,b(z2)|| =∣∣∣∣DT∇h(Dz) +DT b−DT∇h(Dz)−DT b
∣∣∣∣=∣∣∣∣DT (∇h(Dz1)−∇h(Dz2))
∣∣∣∣≤ ||D|| ||∇h(Dz1)−∇h(Dz2)|| .
Thus, to finish the proof for Lipschitz continuity of∇φA,b(x, y), it suffices to show that ||∇h(Dz1)−∇h(Dz2)|| ≤
L ||z1 − z2|| for some constant L. Recalling from the definition of h in (1.3), we see that
∇h(u) = tanh(u
2
):=(
tanh(u(1)
2
), . . . , tanh
(u(k)2
))T,∀u ∈ Rk,∀k (1.11)
7
where we allow the fucntion tanh to be applied component wisely. It follows that
||∇h(Dz1)−∇h(Dz2)|| =∣∣∣∣∣∣∣∣tanh(
Dz12
)− tanh(Dz2
2)
∣∣∣∣∣∣∣∣ .Since tanh is a Lipschitz continuous function with L = 1, we conclude that
∣∣∣∣∣∣∣∣tanh(Dz1
2)− tanh(
Dz22
)
∣∣∣∣∣∣∣∣ ≤ ∣∣∣∣∣∣∣∣D2 (z1 − z2)
∣∣∣∣∣∣∣∣ ≤ ||D||2||z1 − z2||
and consequently,
||∇φA,b(z1)−∇φA,b(z2)|| ≤ ||D||2
2||z1 − z2|| ≤ ||D||2 ||z1 − z2|| .
Thus, ∇φA,b is Lipschitz continuous with Lipschitz constant ||D||2 .
8
Chapter 2
Solving the Binary Logistic
Regression Problem
Let us now take a step back and tackle a larger class of problems, namely smooth convex
optimization. Specifically, consider an optimization problem of the form
x∗ := argminx∈X
f(x) (P)
where X ⊂ Rn is a convex set and f : X → R is a convex function whose gradient satisfies (1.7)
with Lipschitz constant L > 0. In later sections of this chapter, we will discuss methods for solving
(P) along with convergence analysis and limitations of such a scheme. Recall that in Theorem 1.7,
we showed that the function φA,b defined in (1.4) satisfies these conditions. Thus, by solving (P), we
will also develop a method of solving (BLR). Here, we provide an algorithm capable of computing
an ε-approximate solution x such that f(x)− f∗ ≤ ε in O(1)(1/√ε) iterations based on the original
work presented in [1].
2.1 A First Order Method for Smooth Convex Optimization
We now present Nesterov’s accelerated gradient descent for solving (P).
9
Algorithm 1 Nesterov’s accelerated gradient descent (NAGD)
Select parameters γk ∈ (0, 1], ηk. Choose x0 ∈ X. Set y0 = x0.
for k = 1, . . . , N do
zk =(1− γk)yk−1 + γkxk−1 (2.1)
xk =argminx∈X
〈∇f(zk), x〉+ηk2||xk−1 − x||22 (2.2)
yk =(1− γk)yk−1 + γkxk (2.3)
end for
Output yN .
Note that in order to use Algorithm 1 to solve (P), we only require function and gradient
evaluations, an initial feasible point, and constants γk and ηk. We will see in the sequel that, for
different choices of γk, ηk, Algorithm 1 may have different performances.
2.2 Convergence Analysis of Accelerated Gradient Descent
In the following theorem, we state the convergence result of the accelerated gradient descent
in Algorithm 1.
Theorem 2.1. Suppose that we apply Algorithm 1 to solve (P) with parameter γk ∈ [0, 1]. Then
the k-th iterates satisfy
f(yk)− (1− γk)f(yk−1)− γkf(x) ≤ γkηk(||x− xk−1||2 − ||x− xk||2
)+Lγ2k − ηkγk
2||xk − xk−1||2 .
(2.4)
Proof. Let {zi}ki=0, {xi}ki=0 and {yi}ki=0 be the iterate sequences generated by Algorithm 1. Then
by Lipschitz continuity of gradients and convexity, we have the following:
f(yk) ≤ f(zk) + 〈∇f(zk), yk − zk〉+L
2||yk − zk||2 (2.5)
f(yk−1) ≥ f(zk) + 〈∇f(zk), yk−1 − zk〉 (2.6)
f(x) ≥ f(zk) + 〈∇f(zk), x− zk〉. (2.7)
10
Each inequality (2.5), (2.6), and (2.7) follows immediately from (1.8). With these inequalities in
hand, we write
f(yk)− (1− γk)f(yk−1)− γkf(x) ≤ f(zk) + 〈∇f(zk), yk − zk〉+L
2||yk − zk||2
− (1− γk)(f(zk) + 〈∇f(zk), yk−1 − zk〉)
− γk(f(zk) + 〈∇f(zk), x− zk〉).
After simplifying and noting yk − zk = γk(xk − xk−1) from (2.3) and (2.1), we have
f(yk)− (1− γk)f(yk−1)− γkf(x) ≤ 〈∇f(zk), yk − (1− γk)yk−1 − γkx〉+Lγ2k
2||xk − xk−1||2
= 〈∇f(zk), yk − (1− γk)yk−1 − γkxk + γkxk − γkx〉+Lγ2k
2||xk − xk−1||2
= γk〈∇f(zk), xk − x〉+Lγ2k
2||xk − xk−1||2 .
Here, we make use of the definition of yk in (2.3) in the final step. Enforcing the optimality conditions
of xk in (2.2) to γk〈∇f(zk), xk − x〉, we obtain
f(yk)− (1− γk)f(yk−1)− γkf(x) ≤ γkηk(||x− xk−1||2 − ||x− xk||2
)+Lγ2k − ηkγk
2||xk − xk−1||2
which completes our proof.
With Theorem 2.1 in hand, we state two corollaries specifying choices of parameters γk and
ηk.
Corollary 2.2. If we set
γk ≡ 1 and ηk ≡ L (2.8)
in Algorithm 1, then
f(yN )− f(x∗) ≤ L ||x∗ − x0||2
N + 1(2.9)
where yN =∑N
k=0 yk/(N + 1).
Proof. Note that by convexity of f ,
f(yN ) = f
(∑Nk=0 ykN + 1
)≤ 1
N + 1
N∑k=0
f(yk). (2.10)
11
To prove the corollary, it suffices to show that
1
N + 1
(N∑
k=0
f(yk)
)− f(x∗) ≤ L ||x∗ − x0||2
N + 1.
Following the result of Theorem 2.1 with γk ≡ 1 and ηk ≡ L, (2.4) becomes
f(yk)− f(x) ≤ L(||x− xk−1||2 − ||x− xk||2
)(2.11)
Summing (2.11) over all k, the right hand side becomes a telescoping sum to yield
(N∑
k=0
f(yk)
)− (N + 1)f(x) ≤ L
(||x− x0||2 − ||x− xN ||2
)≤ L ||x− x0||2 .
Setting x = x∗ and dividing by (N + 1) gives the desired result.
The parameter setting in (2.8) is desirable in part due to its simplicity. Indeed, under these
settings, xk = yk = zk and
xk = argminx∈X
〈∇f(xk), x〉+L
2||xk−1 − x||2
= argminx∈X
L
2
∣∣∣∣∣∣∣∣x− (xk−1 − 1
L∇f(xk−1)
)∣∣∣∣∣∣∣∣2= argmin
x∈X
∣∣∣∣∣∣∣∣x− (xk−1 − 1
L∇f(xk−1)
)∣∣∣∣∣∣∣∣2
We may understand the above computation of xk as a step in the negative direction of ∇f(xk−1)
projected back into the set X. This technique is referred to as Projected Gradient Descent and is
presented in Algorithm 2.
Algorithm 2 Projected Gradient Descent
Choose x0 ∈ X.
for k = 1, . . . , N do
xk = argminx∈X
∣∣∣∣∣∣∣∣x− (xk−1 −1
L∇f(xk−1)
∣∣∣∣∣∣∣∣2
end for
Output xN =∑N
k=0 xk/(N + 1).
12
Corollary 2.3. If we set
γk =2
k + 1and ηk =
2L
k(2.12)
in Algorithm 1, then
f(yN )− f(x∗) ≤ 4L
N(N + 1)||x∗ − x0||2 . (2.13)
Proof. Define
Γk =
1 k = 1
(1− γk)Γk−1 k > 1
(2.14)
Under the setting in (2.12), it is easy to verify that Γk = 2k(k+1) . Letting x = x∗ in (2.4) and dividing
by Γk gives
1
Γk(f(yk)− f(x∗)) ≤ 1− γk
Γk(f(yk−1)− f(x∗)) +
γkηkΓk
(||x∗ − xk−1||2 − ||xk − x∗||2
).
Subtracting the first term on the right hand side from each side and summing over k gives us another
telescoping series on the left that evaluates to
f(y1)−f(x∗)+
N∑k=2
1
Γk(f(yk)−f(x∗))− 1
Γk−1(f(yk−1)−f(x∗)) ≤
N∑k=1
γkηkΓk
(||x∗ − xk−1||2−||xk − x∗||2)
(2.15)
since (1− γ1)/Γ1 = 0. Evaluating the telescoping series on the, we obtain
1
ΓN(f(yN )− f(x∗)) ≤
N∑k=1
γkηkΓk
(||x∗ − xk−1||2 − ||xk − x∗||2). (2.16)
Applying the parameters (2.12) and (2.14) gives
N(N + 1)
2(f(yN )− f(x∗)) ≤
N∑k=1
4Lk(k + 1)
2k(k + 1)(||x∗ − xk−1||2 − ||xk − x∗||2)
that simplifies to
f(yN )− f(x∗) ≤ 4L
N(N + 1)||x∗ − x0||2 .
Recall that from (2.9), in order to compute an approximate solution y such that f(y) −
13
f(x∗) ≤ ε with parameters defined in (2.8), the number of iterations required is bounded by
O(L||x∗−x0||2ε ). However, under (2.12), (2.13) shows that number is bounded by O
(√4L||x∗−x0||2
ε
).
While the former reduces to the simple scheme in Algorithm 2, the latter gives accelerated conver-
gence and is commonly referred to as Nesterov’s optimal gradient method (OGD).
14
Chapter 3
Optimality of Nesterov’s Optimal
Gradient Method
Algorithm 1 is an example of a first order iterative method. We define a first order method
M as a scheme for solving (P) such thatM initializes the search point x0, andM accesses the first
order information of f through a deterministic oracleOf : Rn → Rn×Rn withOf (x) = (f(x),∇f(x))
for x ∈ Rn. In particular,M can be described by a problem independent x0 and a sequence of rules
{I}∞k=0 such that
xk+1 = I(Of (x0), . . . ,Of (xk)).
Without loss of generality, we may assume that x0 = 0 and that at the N -th iteration, the output
of M is always xN .
In [1], it is shown that the lower complexity bound for smooth convex optimization is
O(1)(1/√ε). In view of Corollary 2.3, we conclude that OGD is an optimal algorithm for smooth
convex optimization, in the sense that its theoretical computational performance can not be im-
proved. For completeness, we present a proof in this chapter.
15
3.1 Iteration Complexity Lower Bound for Smooth Convex
Optimization
We will consider a quadratic programming problem, which falls under the class of smooth
convex optimization, of the form
f∗ := minx∈Rn
f(x) := QA,b(x) :=1
2xTAx− bTx (QP)
where A � 0. Convexity of (QP) follows immediately from Corollary (1.2) since ∇2f(x) = A � 0
by assumption. Furthermore, for any U ∈ Ub where
Ub := {U ∈ Rn×n | Ub = UT b = b and U is orthogonal} (3.1)
we can show that by defining y = Ux for any x ∈ Rn
xTUTAUx = yTAy ≥ 0
follows from A � 0. Thus,
fU (x) := f(Ux) =1
2xTUTAUx− bTx = QUTAU,b(x) (3.2)
is also a smooth convex function and
minx∈Rn
f(x) = minx∈Rn
fU (x).
That is, without loss of generality, we may assume our quadratic program has objective function
fU (x) for some U ∈ Ub. Let Kr(A, b) be the Krylov subspace of order r defined by
Kr(A, b) := span{b, Ab, . . . , Ar−1b}. (3.3)
We have the following lemmas.
Lemma 3.1. For any A ∈ Rn×n, b ∈ Rn, we have the following
16
1. Kr(A, b) ⊆ Kr+1(A, b) for any integer r ≥ 1. Additionally, if Kr(A, b) = Kr+1(A, b), then
Kr(A, b) = Ks(A, b) for all integers s ≥ r.
2. For all U ∈ Ub,
Kr(UTAU, b) = UTKr(A, b). (3.4)
Proof. The subset Kr(A, b) ⊆ Kr+1(A, b) is clear from the definition in (3.3). For the remainder of
the first part, assume that Kr(A, b) = Kr+1(A, b). That is, there must exist ci ∈ R such that
Arb =
r∑i=1
ciAi−1b
This implies that
Ar+1b = A(Arb) = A
r∑i=1
ciAi−1b =
r∑i=1
ciAib =
r−1∑i=1
ciAib+ crA
rb.
The term∑r−1
i=1 ciAib lies in Kr(A, b) by definition whereas crA
rb ∈ Kr+1(A, b) = Kr(A, b) by
assumption. Thus, Ar+1b ∈ Kr(A, b). Inducting on r gives the desired result. For the second
statement, note from (3.1) that
(UTAU)ib = UTAiUb = UTAib.
Thus, x =∑r
i=1 ci(UTAU)i−1b = UT
∑ri=1 ciA
i−1b which implies that Kr(UTAU, b) = UTKr(A, b).
The following lemma appeared in [2] and will be used in our analysis in the sequel.
Lemma 3.2. Let X and Y be two linear subspaces satisfying X ( Y ⊆ Rp. Then for any x ∈ Rp,
there exists orthogonal matrix V such that
V x ∈ Y and V x = x, ∀x ∈ X
Proof. When x ∈ X, V = I satisfies the conditions. Otherwise, let x = y + z where z ∈ X, y ∈ X⊥.
Let {ui}ti=1, t < p be an orthonormal basis for X and extend it to an orthonormal basis {ui}si=1, t <
17
s ≤ p for Y . Define an orthogonal matrix V such that
V ui = ui and V y = ||y||ut+1.
Then for any x ∈ X, there exists constants λi such that∑t
i=1 λiui = x and consequently
V x =
t∑i=1
λiV ui =
t∑i=1
λiui = x.
Furthermore, it follows directly from construction that
V x = z + ||y||ut+1 ∈ Y.
The following two propositions will be crucial to development of the lower complexity bound
for smooth convex optimization.
Proposition 3.1. Let f(x) be defined as in (QP) and N an integer such that
K2N−1(A, b) ( K2N (A, b) ( K2N+1(A, b). (3.5)
Then for any first order scheme M and iterate number k, there exists Uk ∈ Ub such that when M
is applied to minimize fUk(x), the first k iterates satisfy
xi ∈ UTk K2k+1(A, b)
for all i = 0, . . . , k.
Proof. We proceed by induction. When k = 0, the result is trivial since for any U0 ∈ Ub, we have
x0 = 0 ∈ U0 span{b}. Now assume the statement holds for k− 1 < N and let xk be the next iterate
computed by M. By the assumption k − 1 < N and (3.5),
K2k−1(A, b) ( K2k(A, b) ( K2k+1(A, b).
18
and consequently
UTk−1K2k−1(A, b) ( UT
k−1K2k(A, b) ( UTk−1K2k+1(A, b).
by orthogonality of Uk−1. Suppose that UTk−1K2k(A, b) is spanned by {w1, . . . , wl} for some l ≤ 2k.
Then there must exist wl+1 ∈ UTk−1K2k+1(A, b) for which wl+1 6∈ UT
k−1K2k(A, b). By Lemma 3.2 we
can define an orthogonal matrix V satisfying
V wi = wi for wi ∈ UTk−1K2k(A, b) and V vk/ ||vk|| = wl+1. (3.6)
By this definition, it follows that xk ∈ V TUTk−1K2k+1(A, b). We claim that Uk := Uk−1V is the
orthogonal matrix we seek to complete the induction step. It is easy to see that Uk is orthogonal
and since b ∈ UTk−1K2k(A, b),
Ukb = Uk−1V b = Uk−1b = b.
It suffices to show that applying M to fUk(x) generates iterates x0, . . . , xk−1. Indeed, note that for
any x ∈ UTk−1K2k−1(A, b), from (3.4) we have that UT
k−1AUk−1x ∈ UTk−1K2k(A, b). Thus
UTk AUkx = V TUT
k−1AUk−1V x = V T (UTk−1AUk−1x) = UT
k−1AUk−1x.
Hence, for any x ∈ UTk−1K2k−1(A, b), the functions fUk
(x) and fUk−1(x) have the same zeroth
and first order information. Since we assume by the induction hypothesis that x0, . . . , xk−1 ∈
UTk−1K2k−1(A, b), applyingM to minimize fUk
(x) will generate x0, . . . , xk. Noting (3.6), we conclude
that x0, . . . , xk ∈ UTk K2k+1(A, b).
Proposition 3.2. For any A ∈ Rn×n, b ∈ Rn, U ∈ Ub, and any integer r,
minx∈Kr(A,b)
QA,b(x)− minx∈Rn
QA,b(x) = minx∈Kr(UTAUb)
QUTAU,b(x)− minx∈Rn
QUTAU,b(x). (3.7)
19
Proof. Since U ∈ Ub, bTx = UTUbTx = UT bTx = bT (Ux). Noting (3.4),
minx∈Kr(A,b)
1
2xTUTAUx− bTx = min
x∈Kr(A,b)
1
2(xTUT )A(Ux)− bTUx
= minx∈UTKr(A,b)
1
2xTAx− bTx
= minx∈Kr(UTAU,b)
1
2xTAx− bTx.
Also,
minx∈Rn
1
2xTUTAUx− bTx = min
x∈Rn
1
2xTUTAUx− bTUx = min
y∈Rn
1
2yTAy − bT y
with y = Ax. Combining these equalities gives the result.
In view of (3.7), we may simply consider an instance of A and b. Indeed, let
A =L
4
A4k+3 0
0 0
∈ Rn×n, b =L
4e1 (3.8)
with A4k+3 a tridiagonal matrix defined as
A4k+3 =
2 −1
−1 2 −1
−1 2 −1
. . .. . .
. . .
−1 2 −1
−1 2
.
In this definition, L is the Lipschitz constant and ei is the i-th standard basis vector of Rn. We
remark that from [3], the eigenvalues of A are of the form λj = 2 − 2 cos(jπ/(4k + 4)) ≤ 4 for all
j = 1, . . . , 4k + 3. Since A is symmetric, it follows that
||A||2 =L
4||A4k+3|| =
L
4max
i|λj | ≤
4L
4= L.
Thus, ∇QA,b(x) = Ax− b is Lipschitz continuous with Lipschitz constant L.
The optimal solution x∗ to minx∈Rn
QA,b(x) is simply the solution to Ax = b and can be verified
20
to be
x∗i =
1− i
4k+4 1 ≤ i ≤ 4k + 3
0 4k + 4 ≤ i ≤ n.
and so,
QA,b(x∗) = min
x∈RnQA,b(x) =
L
8
(−1 +
1
4k + 4
). (3.9)
In the lemma that follows, we show the final necessary computation for proving the lower complexity
bound.
Lemma 3.3. For A, b defined in (3.8), the Krylov subspaces can be written as
Kr(A, b) =
span{e1, . . . , er} 1 ≤ r ≤ 4k + 2
span{e1, . . . , e4k+3} r ≥ 4k + 3,
(3.10)
the minimum value over K2k+1(A, b) satisfies
minx∈K2k+1(A,b)
QA,b(x) ≥ L
8
(−1 +
1
2k + 2
),
and consequently,
minx∈K2k+1(A,b)
QA,b(x)− minx∈Rn
QA,b(x) ≥ 3L ||x∗||2
128(k + 1)2. (3.11)
Proof. Clearly, b ∈ span{e1} and Ae1 = (2,−1, 0, . . . , 0)T ∈ span{e1, e2}. Continuing, we have
Aei ∈ span{ei−1, ei, ei+1} for all 2 ≤ i ≤ 4k + 2 from the tridiagonal structure. Thus, the first
statement follows. Since x(i) = 0 for any i ≥ 2k + 2 whenever x ∈ K2k+1(A, b), it follows that the
error can be bounded as
minx∈K2k+1(A,b)
QA,b(x) = minx∈K2k+1(A,b)
L
4
(1
2xTAx− bTx
)≥ min
z∈R2k+1
L
4
(1
2zTA2k+1z − z(1)
).
Here, A2k+1 is the (2k + 1)× (2k + 1) leftmost submatrix of A and z is the subvector consisting of
the first 2k + 1 elements of x. In a similar computation that produced minx∈Rn
QA,b(x), we see that
minx∈K2k+1(A,b)
QA,b(x) ≥ L
8
(−1 +
1
2k + 2
).
21
Consequently, the error in terms of the objective function value satisfies
minx∈K2k+1(A,b)
QA,b(x)− minx∈Rn
QA,b(x) ≥ L
32(k + 1). (3.12)
Combining (3.12) with
||x∗||2 =
4k+3∑i=1
(1− i
4k + 4
)2
=
(4k+3∑i=1
1
)−
(2
4k + 4
4k+3∑i=1
i
)+
(1
(4k + 4)2
4k+3∑i=1
i2
)
≤ (4k + 3)− 2
4k + 4· (4k + 3)(4k + 4)
2+
1
(4k + 4)2· (4k + 4)3
3
=4(k + 1)
3
we conclude (3.11).
Theorem 3.3. For any first order iterative method M and iterate k ≤ n−34 , there exists some
smooth convex function g : Rn → R with L-Lipschitz gradient such that xk generated byM satisfies
h(xk)− minx∈Rn
h(x) ≥ 3L ||x0 − x∗||2
128(k + 1)2. (3.13)
Proof. Let f : Rn → R be defined as in (QP) with A and b as stated in (3.8). Applying Proposition
3.1, there exists g(x) := fUk(x) whose iterates x0, . . . , xk lies in the subspace UT
k K2k+1(A, b). Thus,
by Proposition 3.2 and Lemma 3.3,
minx∈Kr(UTAUb)
QUTAU,b(x)− minx∈Rn
QUTAU,b(x) = minx∈Kr(A,b)
QA,b(x)− minx∈Rn
QA,b(x)
≥ 3L ||x∗||2
128(k + 1)2.
In view of Theorem 3.3, our proposed Algorithm 1 with parameters (2.12) is an optimal
algorithm for solving smooth convex optimization. Futhermore, there exists the following other
available lower complexity bound results on deterministic first order methods for convex optimization
f∗ := minxf(x).
• When f is convex, the lower complexity bound is O(1)(1/ε2) [4, 1].
22
• When f is convex, nonsmooth with bilinear saddle point structure, the lower complexity bound
is O(1)(1/ε) [2].
• When f is strongly convex, smooth the lower complexity bound is O(1) log(1/ε) [1, 5].
23
Chapter 4
Lower Complexity Bound for
Binary Logistic Regression
Let us now consider our (BLR) again. By (1.7), we know that we can solve (BLR) via
Algorithm 1. Hence, in view of (2.3), an ε− approoximate solution can be computed in O(1)(1/√ε)
first order oracle iterations. However, it has yet to be determined that that Algorithm 1 achieves the
lower complexity bound for binary logistic regression problems via first order deterministic methods.
We will develop a lower complexity bound for functions of the form in (BLR), a subset of smooth
convex optimization. In doing so, we will construct a worst-case dataset for solving binary logistic
regression that requires O(1)(1/√ε) first order oracle calls and thus, showing Algorithm 1 is an
optimal algorithm for this class of problems. These worst-case constructions will satisfy y∗ = 0.
Consequently, it suffices to solve the logistic model with homogeneous linear predictor
lA,b(x) = h(Ax)− bTAx (4.1)
and corresponding problem
l∗A,b = minx∈Rn
lA,b(x). (HBLR)
24
4.1 Linear Span Assumption
In this section, we make the following simplifying assumption: the iterates of a first order
deterministic method M satisfy
xt ∈ span{∇f(x0), . . . ,∇f(xt−1)} (4.2)
for all t ≥ 1. We will refer to (4.2) as the linear span assumption. This assumption is convenient in
demonstrating our desired result, but we will later show that our results can also be proved without
the linear span assumption through a technique developed in [6] (see also [2]).
Define the following for a parameterization of binary logistic regression
Wk :=
−1 1
−1 1
. ..
. ..
−1 1
−1
∈ Rk×k, Ak :=
2σWk
−2ζWk
−2σWk
2ζWk
∈ R4k×k, bk =
12k
−12k
∈ R4k. (4.3)
Note that the above construction follows the idea in [1]; indeed, W 2k appeared in the construction
(3.8). Let us consider performing binary logistic regression with data matrix Ak and response vector
bk. In order to avoid duplicate data entries in Ak, we also assume without loss of generality that
σ > ζ > 0. Denote then the functions
fk(x) := h(Akx)− bTk (Akx) and φk(x, y) := h(Akx+y1k)− bTk (Akx+ y1k). (4.4)
In view of the discussion above, we can view fk and φk as objective functions of homoegeneous and
inhomogeneous binary logistic regression problems respectively. Before we proceed, we note that a
sufficient optimality condition for minimizing the function f in (4.4) is
∇fk(x) = 0. (4.5)
We will use three lemmas to study the behavior of the iterates of a first order method M applied
to (HBLR). For brevity, denote et,k the t-th standard basis in Rk.
25
Lemma 4.1. For Ak defined in (3.8),
||Ak|| ≤√
32(σ2 + ζ2). (4.6)
Proof. Let u ∈ Rk. Then we have
||Wku||2 = (u(k) − u(k−1))2 + · · ·+ (u(2) − u(1))2 + (u(1))2
≤ 2((u(k))
2 + (u(k−1))2 + · · ·+ (u(2))
2 + (u(1))2 + (u(1))
2)
≤ 4 ||u||2 .
This implies that
||Aku||2 = 8(σ2 + ζ2) ||Wku||2 ≤ 32(σ2 + ζ2) ||u||2 .
Thus,
||Ak|| ≤√
32(σ2 + ζ2).
Lemma 4.2. The minimization problem
f∗k := minx∈Rk
fk
has optimal solution
x∗ = c(1, 2, . . . , k)T
with optimal value
f∗k = 8k log 2 + 4k (log cosh(σc) + log cosh(ζc)− (σ − ζ)c)
where c satisfies
σ tanh(σc) + ζ tanh(ζc) = σ − ζ. (4.7)
In addition, (x∗, 0) is the unique optimal solution to minx∈Rn,y∈R φk(x, y).
26
Proof. Since Wkx∗ = c1k follows from the definition in (3.8), it is then easy to check that
∇h(Akx∗) = tanh
1
2
2σc1k
−2ζc1k
−2σc1k
2ζc1k
=
tanh(σc)1k
tanh(σc)1k
tanh(σc)1k
tanh(σc)1k
via the description in (1.11). Using above equation and noting that Wk1k = ek,k, we have
ATk∇h(Akx
∗) = 4(σ tanh(σc) + ζ tanh(ζc))ek,k.
Combining the above with
ATk bk = 4(σ − ζ)ek,k (4.8)
we conclude that the optimality condition in (4.5) is met whenever c satisfies (4.7). Such a c is well-
defined since the function T (c) := σ tanh(σc)+ζ tanh(ζc)−σ+ζ is continuous with T (0) = −σ+ζ < 0
and limc→∞
T (c) = 2ζ > 0. Moreover, noting that cosh is an even function,
f∗k = h(Akx∗)− 4k(σ − ζ)c
= 2k (log(2 cosh(σc)) + log(2 cosh(−ζc)) + log(2 cosh(σc)) + log(2 cosh(ζc)))− 4k(σ − ζ)c
= 8k log 2 + 4k (log cosh(σc) + log cosh(ζc)− (σ − ζ)c) .
Furthermore, after noting
∂
∂y
∣∣∣∣y=0
φk(x∗, y) = 1Tk∇h(Akx∗)− bTk 1k = 0
and
∇xφk(x∗, 0) = ∇fk(x∗) = 0,
we conclude that
27
Lemma 4.3. Let t, k be positive integers satisfying t ≤ k. Denote
Xt,k := span{ek−t+1,k, . . . , ek,k}, ∀k, 1 ≤ t ≤ k
and
Yt,k := span{e1,4k, . . . , et,4k, . . . , ek+1,4k, . . . , ek+t,4k, . . . , e2k+1,4k, . . . , e2k+t,4k, . . . , e3k+1,4k, . . . , e3k+t,4k}.
Then for any x ∈ Xt,k, Akx,∇h(Akx) ∈ Yt,k and ATk∇h(Akx), ∇fk(x) ∈ Xt+1,k. Furthermore,
minx∈Xt,k
fk(x) = 8(k − t) log 2 + minu∈Rt
ft(u). (4.9)
Proof. Let x ∈ Xt,k. Thus, x can be decomposed into x = (0Tk−t, uT )T for u ∈ Rt. Consequently,
Akx =
2σWtu
0k−t
−2ζWtu
0k−t
−2σWtu
0k−t
2ζWtu
0k−t
, ∇h(Akx) =
tanh(σWtu)
0k−t
tanh(−ζWtu)
0k−t
tanh(−σWtu)
0k−t
tanh(ζWtu)
0k−t
.
Hence, Akx,∇h(Akx) ∈ Yt,k. Noting that, since tanh is an odd function, we have
ATk∇h(Akx) = 4σWT
k
tanh(σWtu)
0k−t
+ 4ζWTk
tanh(ζWtu)
0k−t
and consequently
∇fk(x) = ATk∇h(Akx)−AT
k bk ∈ Xt+1,k.
28
To show (4.9) note that, for x ∈ Xt,k,
h(Akx) =
t∑i=1
2 log(2 cosh(σWtu)(i)) + 2 log(2 cosh(−ζWtu)(i))
+ 2 log(2 cosh(−σWtu)(i)) + 2 log(2 cosh(ζWtu)(i)) + 8(k − t) log cosh(0)).
Applying again the definition of Wk in (1.3) here, we have h(Akx) = 8(k−t) log 2+h(Atu). Moreover,
since
bTkAkx = 4(σ − ζ)u = bTt Atu
from the definitions in (3.8), we conclude the result (4.9) immediately.
Similar to the discussion around Proposition 3.1, as a consequence of the previous lemma, we
show that a first order methodM satisfying the linear span assumption generates iterates xt ∈ Xt,k
when minimizing fk(x).
Lemma 4.4. Suppose thatM is a deterministic first order method whose iterates satisfy the linear
span assumption (4.2). When M is applied to minimize fk defined in (4.4), we have xt ∈ Xt,k for
1 ≤ t ≤ k.
Proof. We will proceed via induction. Consider the case t = 1. By (4.2), x1 ∈ span{∇fk(x0)}. Since
x0 = 0, ∇fk(x0) = −ATk bk. Thus, in view of (4.8),
∇fk(x0) ∈ span{ek,k} = X1,k
so the statement holds for t = 1. Continuing with induction, assume that xi ∈ Xi,k for 1 ≤ i ≤ t < k.
Noting lemma 4.2, ∇fk(xi) ∈ Xi+1,k for all t. Thus, by (4.2), it follows that
xs+1 ∈ span{∇fk(x0), . . . ,∇fk(xs)} ⊆ Xt+1,k
which completes the induction process.
29
Combining Lemmas 4.3, 4.4, and 4.2, it follows that
fk(xt)− f∗k ≥ minx∈Xt,k
fk(x)− f∗k (4.10)
= 8(k − t) log 2 + f∗t − f∗k (4.11)
= 4(k − t) ((σ − ζ)c− log cosh(σc)− log cosh(ζc)) . (4.12)
In an attempt to simplify the lower complexity bound presented above, we present the following
lemma.
Lemma 4.5. For any real numbers σ and ζ that satisfy σ/ζ = 1.3,
(σ − ζ)c− log cosh(σc)− log cosh(ζc) ≥ c2σ2
2.
Proof. For c > 0, the function c→ c tanh(c) is increasing since its derivative is positive. Thus, since
ζc < σc,
ζc tanh(ζc) ≤ σc tanh(σc)
and consequently
c tanh(ζc) ≤ c tanh(σc).
Applying these inequalities to bound (4.7) gives
2ζ tanh(ζc) ≤ σ − ζ ≤ 2σ tanh(σc)
Furthermore, because function c 7→ tanh(c) has positive derivative everywhere, we conclude that
c ∈ [a, b] where
a =1
σarctanh
(1
2− ζ
2σ
), b =
1
ζarctanh
(σ
2ζ− 1
2
). (4.13)
Here, since σ/ζ = 1.3, we have a, b > 0. Thus, applying (4.7) we have
(σ−ζ)c−log cosh(σc)−log cosh(ζc) =(cσ)2
(cσ)2(σc tanh(σc) + ζc tanh(ζc)− log cosh(σc)− log cosh(ζc)) .
Noting again by the first derivative that function c 7→ c tanh(c)− log cosh(c) is increasing for c > 0,
30
we may apply (4.13) and obtain
(σ − ζ)c− log cosh(σc)− log cosh(ζc) ≥ c2σ2C (4.14)
where
C :=1
(bσ)2(σa tanh(σa)− log cosh(σa) + ζa tanh(ζa)− log cosh(ζa)) .
Since C only depends on σ/ζ = 1.3, we may numerically verify that
C >1
2(4.15)
and the statement of the lemma follows from (4.14) and (4.15) immediately.
We are now ready to estimate the lower complexity bound of first order methods applied to
binary logistic regression under the linear span assumption.
Theorem 4.1. Let M be a deterministic first order method applied to solve (BLR) whose iterates
satisfy the linear span assumption (4.2). For any iteration count M and constants n = 2M , N = 8M ,
there exist data matrix A ∈ RN×n and response vector b ∈ {−1, 1}N such that the M -th iterate
generated by M satisfies
φA,b(xM )− φA,b(x∗) ≥ 3 ||A||2 ||x0 − x∗||2
24(2M + 1)(4M + 1)(4.16)
and
||xM − x∗||2 >1
8||x0 − x∗||2 . (4.17)
Proof. We begin by setting constants ζ > 0 and σ = 1.3ζ and defining Ak as in (3.8). Let M be
applied to minimize fk defined in (4.4) with k = 2M . Then from Lemma 4.4, the discussion in after
(4.10), and the simplification in Lemma 4.5, we have
fk(xt)− f∗k ≥ 2(k − t)c2σ2, ∀t ≤ k (4.18)
Furthermore, from (4.6) and Lemma 4.2 along with the assumption that x0 = 0, we have
||Ak|| ≤ 4√
2σ2 + 2(σ/1.3)2 < 8σ (4.19)
31
and
||x0 − x∗||2 = c2k∑
i=1
i2 =c2
6k(k + 1)(2k + 1). (4.20)
Combining the above two relations with (4.10), we conclude
fk(xt)− f∗k >3(k − t) ||Ak||2 ||x0 − x∗||2
16k(k + 1)(2k + 1). (4.21)
Continuing, since xt ∈ Xt,k by Lemma 4.4, we have the bound
||xt − x∗||2 ≥ c2k−t∑i=1
i2 =c2
6(k − t)(k − t+ 1)(2k − 2t+ 1). (4.22)
Consequently, by setting t = M and noting k = 2M , (4.21) becomes
fk(xM )− f∗k >3 ||Ak||2 ||x0 − x∗||2
24(2M + 1)(4M + 1).
Applying (4.20) and (4.22) results in
||xM − x∗||2 ≤c2
6M(M + 1)(2M + 1) >
c2
48(2M)(2M + 1)(4M + 1) =
1
8||x0 − x∗||2 .
Letting A := Ak and b := bk, we conclude the theorem.
4.2 Lower Complexity Bound for solving Binary Logistic Re-
gression via General first order Methods
In this section we will remove the assumption in (4.2) and build on the results of Theorem
4.1 to establish a similar result without the linear span assumption.
Lemma 4.6. For Ak, bk specified in (3.8), any first order method M, and some t ≤ k−32 , there
exists an orthogonal matrix Ut ∈ Rk×k satisfying
1. UtATk bk = AT
k bk
2. When M is applied to solve (HBLR), the iterates x0, . . . , xt satisfy
xi ∈ UTt X2i+1,k, i = 0, . . . , t.
32
Proof. Let us first define the set
U := {V ∈ Rk×kV | is orthogonal and V ATk bk = AT
k bk}.
We will proceed by induction. The case t = 0 follows immediately from letting U0 be the identity
matrix. Now suppose the statement holds with t = s−1 < (k−1)/2 and denote xs the next iterate.
We will show the statement holds with t = s. Since s < (k−1)/2 is sufficiently small, it follows that
X2s,k ⊂ X2s+1,k and consequently UTs−1X2s,k ⊂ UT
s−1X2s+1,k. Moreover, by Lemma 3.2, there exists
some orthogonal matrix V such that
V x = x ∀x ∈ UTs−1X2s,k and V xs ∈ UT
s−1X2s+1,k. (4.23)
Denote then
Us = Us−1V. (4.24)
From computation in (4.8), we know that ATk bk ∈ X1,k ⊂ X2s,k. Thus, we have UT
s ATk bk =
V TUTs−1A
Tk bk = V TAT
k bk = ATk bk. Noting that V is the product of orthogonal matrices and thus
orthogonal, we conclude that Us ∈ U . Now suppose x ∈ UTs X2s−1,k. By (4.23) and the fact that
Us ∈ U , we have
lAkUs,bk(x) = h(AkUs)− xTUTs A
Tk bk = h(AkUs−1x)− xTUT
s−1ATk bk = lAkUs−1,bk(x).
Additionally, applying Lemma 4.2 and (4.23), we see that V TUTs−1A
Tk∇h(AkUs−1x) = UT
s−1ATk∇h(AkUs−1x)
and therefore,
∇lAkUs,bk(x) = UTs A
Tk∇h(AkUsx)− UT
s ATk bk = UT
s−1ATk∇h(AkUsx)− UT
s−1ATk bk = ∇lAkUs−1,bk(x).
Therefore, for any x ∈ UTs X2s−1,k, the first order method M receives same information regardless
whether it is applied to lAkUs,bk or lAkUs−1,bk . Thus, it produces the same iterates x0, . . . , xs−1 when
minimizing lAkUs,bk as it does when minimizing lAkUs−1,bk . Furthermore, by the construction of V ,
for any i = 0, . . . , s− 1, we have
UTs X2i+1,k = V TUT
s−1X2i+1,k = UTs−1X2i+1,k (4.25)
33
and
xs ∈ V TUTs−1X2i+1,k = UT
s X2s+1,k.
Combining these two arguments with the induction hypothesis yields
xi ∈ UTs X2i+1,k, ∀i = 0, . . . , s. (4.26)
In view of (4.26), we conclude the induction for the t = s case by choosing U = Us.
Theorem 4.2. For any first order method M and fixed iteration number M with corresponding
constants N = 10M+8, n = 4M+2, there always exists data matrix A ∈ RN×n and response vector
b ∈ RN such that when M is applied to solve (HBLR), the M -th iterate satisfies
lA,b(xM )− l∗A,b ≥3 ||A||2 ||x0 − x∗||2
16(4M + 3)(8M + 5)
and
||xM − z∗||2 >1
8||x0 − z∗||2
where z∗ is the minimizer of lA,b.
Proof. Let ζ > 0 and set σ = 1.3ζ. Let k = 4M + 2 and define Ak using σ, ζ as in (3.8). Lemma
4.6 provides an orthogonal matrix U such that UTATk bk = AT
k bk and when M is applied to solve
lAkU,bk , the iterates xi satsify xi ∈ UTX2i+1,k for any 0 ≤ i ≤M . Thus, we have
lAkU,bk(xM ) ≥ minx∈UTX2M+1,k
lAkU,bk(x)
= minx∈UTX2M+1,k
h(AkUx)− xTUTATk bk
= minx∈X2M+1,k
h(Akx)− xTATk bk
= minx∈X2M+1,k
fk(x)
34
and
l∗AkU,bk= min
x∈RklAkU,bk(x)
= minx∈Rk
h(AkUx)− xTUTATk bk
= minx∈Rk
h(Akx)− xTATk bk
= minx∈Rk
fk(x).
Here, we used the definition of fk as in (4.4). Noting the two above computations applying (4.10)
and Lemma 4.5, we see that
lAkU,bk(xT )− l∗AkU,bk≥ min
x∈X2M+1,k
fk(x)− minx∈X2M+1,k
fk(x)
= 4(k − 2M − 1)((σ − ζ)c− log cosh(σc)− log cosh(ζc))
≥ 2(k − 2M − 1)c2σ2.
(4.27)
In view of computation of l∗AkU,bk, we see that its minimizer z∗ must satisfy zT = UTx∗ with x∗
being the minimizer of fk provided in Lemma 4.2. Since Lemma 4.6 guarantees xM ∈ UTX2M+1,k,
it follows that
||xM − z∗|| ≥ maxx∈X2M+1,k
||x− x∗||2
≥ c2k−2M−1∑
i=1
i2
≥ c22M+1∑i=1
i2
=c2
6(2M + 1)(2M + 2)(4M + 3)
since we set k = 4M + 2. Furthermore, because x0 = 0,
||x0 − z∗||2 =∣∣∣∣UTx∗
∣∣∣∣2 = ||x∗||2 = c24M+2∑i=1
i2 =c2
6(4M + 2)(4M + 3)(8M + 5) (4.28)
35
and consequently
||xM − z∗||2 >1
8||x0 − z∗|| . (4.29)
Lastly, applying (4.28) to (4.27) and noting k = 4M + 2, we see that
lAkU,bk(xM )− l∗AkU,bk≥ 6σ2(2M + 1) ||x0 − z∗||2
(2M + 1)(4M + 3)(8M + 5).
Recalling (4.6) we conclude that
lAkU,bk(xM )−l∗AkU,bk≥ 6σ2 ||A||2 ||x0 − z∗||2
||A||2 (4M + 3)(8M + 5)≥ 6σ2 ||A||2 ||x0 − z∗||2
32σ2(4M + 3)(8M + 5)=
3 ||A||2 ||x0 − z∗||2
16(4M + 3)(8M + 5).
along with (4.29) completes the proof by setting A := AkU and b := bk.
4.3 Conclusions
With Theorem 4.2, we are able to construct a homogeneous worst case dataset for binary
logistic regression which proves that Algorithm 1 is an optimal first order method for such a class of
functions. However, the dataset generated here is most definitely non-standard. Possible extensions
of this work would be to substitute our contrived Ak, bk in favor of randomly generated datasets for
more practical results.
36
Bibliography
[1] Y. E. Nesterov. Introductory Lectures on Convex Optimization: A Basic Course. Kluwer Aca-demic Publishers, Massachusetts, 2004.
[2] Yuyuan Ouyang and Yangyang Xu. Lower complexity bounds of first-order methods for convex-concave bilinear saddle-point problems. Mathematical Programming, pages 1–35, 2019.
[3] L. Pasquini S. Noschese and L. Reichel. Tridiagonal toeplitz matrices: properties and novelapplications. Numerical Linear Algebra with Applications, 2012.
[4] A. Nemirovski and D. Yudin. Problem complexity and method efficiency in optimization. Wiley-Interscience Series in Discrete Mathematics. John Wiley, XV, 1983.
[5] Blake Woodworth, Jialei Wang, Brendan McMahan, and Nathan Srebro. Graph oracle models,lower bounds, and gaps for parallel stochastic optimization. arXiv preprint arXiv:1805.10222,2018.
[6] A. S. Nemirovski. Information-based complexity of linear operator equations. Journal of Com-plexity, 8(2):153–175, 1992.
37